Quantum machine learning for recognition of defects in ultrasonic imaging

Abstract

The paper discusses a new paradigm of employing a quantum machine learning (QML) algorithm for automated weld defect recognition. A variational quantum classifier (VQC) using ultrasonic phased arrays is proposed to extract weld defect features in the atomic state to improve the classification accuracy and achieve high-speed calculation due to simultaneous qubits. The VQC is trained using a simulation-assisted weld dataset generated using finite element (FE) models and deep convolution generative adversarial networks (DCGAN). The total focusing method (TFM) weld images of porosity and slag are generated using time-transmitted signals received by performing full matrix capture, modeling various defect morphologies using FE simulations. These datasets are fed to train the DCGAN to generate synthetic TFM images. We use the feature selection method to obtain the best results with a quantum circuit with minimal qubits. Prominent features so obtained are supplied to the encoder circuit of the VQC to convert it to a quantum domain, thereby passing to an ansatz circuit to train quantum hyperparameters. The loss is computed for every iteration by optimizing the learnable parameters of the VQC. The VQC is trained by varying quantities of datasets to improve the reliability and efficiency of the weld defect classifications. It is found that VQC outperforms some of the classical machine learning algorithms with an accuracy of $96$%.